Downloads · 30 days
0
EdwardoSunny/ttt-discover-v3-qwen3-8b-erdos-s2-step49
ttt-discover-v3-qwen3-8b-erdos-s2-step49 is a machine learning model from EdwardoSunny. Use it for the machine learning task on the model card, and read the license before you ship it in a product.
LoRA adapter (rank 32, all-linear) from the upstream-faithful local reproduction of TTT-Discover on autoresearch-bench erdos. Checkpoint saved after training step 48 (0-indexed). Strict upstream eval parity: 1100s har…
Downloads · 30 days
0
Access
Public
Updated Jul 11, 2026
Repo size
349 MB
Likes
0
Public
Click a slice to open those files.
.safetensors349 MB · 100%
From the Hugging Face model README
LoRA adapter (rank 32, all-linear) from the upstream-faithful local
reproduction of TTT-Discover on autoresearch-bench erdos. Checkpoint saved
after training step 48 (0-indexed). Strict upstream eval parity:
1100s hard kill, verbatim prompts/entrypoints, group 64x8, T=1.0, kl 0.1.
{
"step": 48,
"progress/batch": 48,
"optim/lr": 4e-05,
"progress/done_frac": 0.98,
"puct/buffer_size": 776,
"puct/sampled_size": 8,
"puct/T": 24576,
"puct/scale_last": 0.2791161009494976,
"puct/buffer_value/mean": -0.38404527751687056,
"puct/buffer_value/std": 0.018881292036034177,
"puct/buffer_value/min": -0.66,
"puct/buffer_value/max": -0.3808838990505024,
"puct/buffer_timestep/mean": 23.24742268041237,
"puct/buffer_timestep/std": 14.002231727743395,
"puct/buffer_timestep/min": -1.0,
"puct/buffer_timestep/max": 47.0,
"puct/buffer_construction_len/mean": 112.6881443298969,
"puct/buffer_construction_len/std": 55.56454094077041,
"puct/buffer_construction_len/min": 40.0,
"puct/buffer_construction_len/max": 400.0,
"puct/sampled_value/mean": -0.3808843215075235,
"puct/sampled_value/std": 4.235131738593847e-07,
"puct/sampled_value/min": -0.3808848037470758,
"puct/sampled_value/max": -0.3808838990505024,
"puct/sampled_timestep/mean": 47.0,
"puct/sampled_timestep/std": 0.0,
"puct/sampled_timestep/min": 47.0,
"puct/sampled_timestep/max": 47.0,
"puct/sampled_construction_len/mean": 200.0,
"puct/sampled_construction_len/std": 0.0,
"puct/sampled_construction_len/min": 200.0,
"puct/sampled_construction_len/max": 200.0,
"time/sampling": 2080.3393607139587,
"env/all/ac_tokens_per_turn": 9906.302734375,
"env/all/ob_tokens_per_turn": 1862.625,
"env/all/turns_per_episode": 1.0,
"env/all/total_episodes": 512,
"env/all/total_turns": 512,
"env/all/total_ac_tokens": 5072027,
"env/all/total_ob_tokens": 953664,
"env/all/time/sampling_mean": 588.4042912237346,
"env/all/time/sampling_max": 851.4134395122528,
"env/all/time/env_step_mean": 187.20139834657311,
"env/all/time/env_step_max": 1248.5352103710175,
"env/all/reward/mean": 0.8630208356639941,
"env/all/reward/max": 2.6254735914739795,
"env/all/reward/min": 0.0,
"env/all/format": 1.0,
"env/all/format/min": 1.0,
"env/all/format/max": 1.0,
"env/all/reward": 0.8630208356639941,
"env/all/correctness": 0.3515625,
"env/all/correctness/min": 0.0,
"env/all/correctness/max": 1.0,
"env/all/raw_score": 0.41364382303070024,
"env/all/raw_score/min": 0.3808836535216678,
"env/all/raw_score/max": 0.664988109648523,
"env/all/initial_raw_score": -0.38088432150752344,
"env/all/initial_raw_score/min": -0.3808848037470758,
"env/all/initial_raw_score/max": -0.3808838990505024,
"env/all/msg": "RuntimeError: Program execution failed: ValueError: The truth value of an array with more than one element is ambiguous. Use a.any() or a.all()",
"env/all/parsed_code": "```python\nimport numpy as np\nfrom scipy.optimize import minimize\n\ndef run(seed=42, budget_s=1000, **kwargs):\n np.random.seed(seed)\n\n n_points = 200 # Balances precision and speed\n dx = 2.0 / n_points\n desired_sum = n_points / 2.0 # \u222bh(x) dx = 1\n\n best_h = np.zeros(n_points)\n best_t = np.inf\n\n # Check if initial_h_values is available\n initial_h_values = globals().get('initial_h_values', [])\n if initial_h_values and len(initial_h_values) == n_points:\n initial_h = np.array(initial_h_values)\n else:\n initial_h = np.zeros(n_points)\n for i in range(n_points):\n if i % 2 == 0:\n initial_h[i] = 1.0 # Basic periodic initialization\n\n # Generate diverse initial guesses\n initial_guesses = [initial_h.copy()]\n\n # Add perturbations for diversity\n for _ in range(4): # 4 random variations\n noise = np.random.randn(n_points) * 0.1\n noise = np.clip(noise, -0.2, 0.2)\n noisy_h = initial_h + noise\n noisy_h = np.clip(noisy_h, 0, 1)\n \n # Adjust sum\n sum_h = np.sum(noisy_h)\n delta = desired_sum - sum_h\n for i in reversed(range(n_points)):\n if delta <= 0:\n break\n add = min(1.0 - noisy_h[i], delta)\n noisy_h[i] += add\n delta -= add\n \n initial_guesses.append(noisy_h.copy())\n\n for h_guess in initial_guesses:\n h_guess = h_guess.copy()\n sum_h = np.sum(h_guess)\n delta = desired_sum - sum_h\n for i in reversed(range(n_points)):\n if delta <= 0:\n break\n add = min(1.0 - h_guess[i], delta)\n h_guess[i] += add\n delta -= add\n h_guess = np.clip(h_guess, 0, 1)\n h_vec = h_guess\n\n # Compute cross-correlation and select relevant shifts\n cross_corr = np.correlate(h_vec, 1 - h_vec, mode='full') * dx\n indices = [n_points - 1 - 2, n_points - 1 - 1, n_points - 1, n_points - 1 + 1, n_points - 1 + 2]\n valid_indices = [i for i in indices if 0 <= i < len(cross_corr)]\n selected_overlaps = cross_corr[valid_indices]\n initial_t = np.max(selected_overlaps) if selected_overlaps.size > 0 else 0.5\n\n # Optimization function\n def objective_func(x):\n return x[-1] # Minimize t\n\n def constraint_sum(x):\n return np.sum(x[:-1]) - desired_sum\n\n # Constraint function for relevant shifts\n def constraints_func(x):\n h_vec = x[:-1]\n t = x[-1]\n cross_corr = np.correlate(h_vec, 1 - h_vec, mode='full') * dx\n valid_indices = [n_points - 1 - 2, n_points - 1 - 1, n_points - 1, n_points - 1 + 1, n_points - 1 + 2]\n valid_indices = [i for i in valid_indices if 0 <= i < len(cross_corr)]\n selected_overlaps = cross_corr[valid_indices]\n return [t - selected_overlaps[i] for i in range(len(valid_indices))] # Ensure t >= selected overlaps\n\n # Bounds and constraints\n bounds = [(0.0, 1.0) for _ in range(n_points)] + [(0.0, np.inf)]\n cons = [\n {'type': 'eq', 'fun': constraint_sum},\n {'type': 'ineq', 'fun': constraints_func}\n ]\n\n # Optimize with SLSQP\n try:\n result = minimize(\n fun=objective_func,\n x0=np.concatenate((h_guess, [initial_t])),\n bounds=bounds,\n constraints=cons,\n method='SLSQP',\n options={\n 'ftol': 1e-8,\n 'gtol': 1e-6,\n 'maxiter': 2000,\n 'disp': False\n }\n )\n if result.success:\n current_t = result.x[-1]\n if current_t < best_t:\n best_t = current_t\n best_h = result.x[:-1]\n except:\n pass\n\n return best_h, best_t, n_points\n```",
"env/all/time/policy": 588.4042912237346,
"env/all/time/policy/min": 176.7600018978119,
"env/all/time/policy/max": 851.4134395122528,
"env/all/time/env_step": 187.20139834657311,
"env/all/time/env_step/min": 0.0069522857666015625,
"env/all/time/env_step/max": 1248.5352103710175,
"env/all/time/reward_compute": 3.6135315895080566e-07,
"env/all/time/reward_compute/min": 3.129243850708008e-07,
"env/all/time/reward_compute/max": 4.023313522338867e-07,
"env/all/by_group/frac_mixed": 1.0,
"env/all/by_group/frac_all_good": 0.0,
"env/all/by_group/frac_all_bad": 0.0,
"advantage/mean": 0.02621169202029705,
"advantage/min": -0.9999990463256836,
"advantage/max": 5.342919826507568,
"time/assemble_training_data": 9.397233486175537,
"time/kl_vs_base": 143.46742224693298,
"kl_policy_base": 0.0008664666675031185,
"time/train": 1116.0691933631897,
"time/save_checkpoint": 7.9501588344573975,
"time/total": 3359.300264596939
}
[2026-07-09T06:24:18+00:00] job=1812626 node=node-31 ngpu=3 ntrain=1 replicas=2 flash_attn=no
[2026-07-09T08:02:36+00:00] job=1813125 node=node-12 ngpu=3 ntrain=1 replicas=2 flash_attn=yes
[2026-07-09T09:32:13+00:00] job=1813126 node=node-3 ngpu=6 ntrain=2 replicas=4 flash_attn=yes
[2026-07-09T10:04:31+00:00] job=1813623 node=node-31 ngpu=3 ntrain=1 replicas=2 flash_attn=yes